A feature that survives scale and rotation

In September 1999 David Lowe presented SIFT: stable points are found as extrema in a difference-of-Gaussians pyramid, and the neighbourhood of each is described by a vector of gradient orientation histograms. Over twenty images and about 15,000 keys, 85.4 percent of keys were still found after a 20 degree rotation, 85.1 percent after a scaling by 0.7, and 90.3 percent after adding 10 percent pixel noise.

Features are described by vectors of orientation planes. In the paper's implementation that is 8 planes over a 4 by 4 grid at one pyramid level plus 8 planes over a 2 by 2 grid an octave higher — 160 samples in total. One image yields on the order of a thousand keys in under a second, and recognising an object in a cluttered scene takes under two seconds. The paper calls recognition of planar objects robust to at least a 60 degree rotation of the plane away from the camera. The record does not claim a "128-dimensional descriptor": in this paper the vector holds 160 samples, and 128 appears in the later journal version of the method. Nor does it claim "over 90 percent correct matches at up to a 60 degree rotation": 90.3 percent is the top row of the stability table and it is about 10 percent pixel noise, while the 60 degrees concern object recognition rather than a fraction of matches. US patent 6,711,293 on the method carries a priority date of 8 March 1999 — earlier than the conference — was filed on 6 March 2000 and granted on 23 March 2004 to the University of British Columbia.